Accurate Many-Body Repulsive Potentials for Density-Functional Tight Binding from Deep Tensor Neural Networks.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Materials Science

Background:

• Density-functional tight binding (DFTB) offers a computationally efficient approach for electronic structure calculations.
• Deep tensor neural networks (DTNN) have shown promise in modeling complex molecular interactions.
• Accurate prediction of molecular properties is crucial for drug discovery and materials design.

Purpose of the Study:

• To develop a hybrid model combining DFTB and DTNN for enhanced molecular property prediction.
• To improve the modeling of localized many-body interatomic repulsive energy.
• To achieve chemically accurate predictions for structural, energetic, and vibrational properties of organic molecules.

Main Methods:

• Integration of deep tensor neural networks (DTNN) into density-functional tight binding (DFTB) framework.
• Development of a nonlinear model for localized many-body interatomic repulsive energy.
• Validation against hybrid DFT-PBE0 functional for various molecular properties.

Main Results:

• The novel DFTB-NN_rep model significantly improves upon standard DFTB and DTNN.
• Accurate predictions of atomization energies, isomerization energies, equilibrium geometries, and vibrational frequencies.
• Reliable prediction of dihedral rotation profiles for diverse organic molecules.

Conclusions:

• The DFTB-NN_rep approach demonstrates the potential of combining semiempirical methods with machine learning.
• This hybrid model accurately captures localized many-body interactions.
• Future advancements could enable accurate electronic structure calculations for large-scale systems.